Northstar enables automatic classification of known and novel cell types from tumor samples

Abstract

Single cell transcriptomics is revolutionising our understanding of tissue and disease heterogeneity, yet cell type identification remains a partially manual task. Published algorithms for automatic cell annotation are limited to known cell types and fail to capture novel populations, especially cancer cells. We developed northstar, a computational approach to classify thousands of cells based on published data within seconds while simultaneously identifying and highlighting new cell states such as malignancies. We tested northstar on data from glioblastoma, melanoma, and seven different healthy tissues and obtained high accuracy and robustness. We collected eleven pancreatic tumors and identified three shared and five private neoplastic cell populations, offering insight into the origins of neuroendocrine and exocrine tumors. Northstar is a useful tool to assign known and novel cell type and states in the age of cell atlases.

Disruption of a GATA2-TAL1-ERG regulatory circuit promotes erythroid transition in healthy and leukemic stem cells

“…Aligned sequencing data was displayed in BigWig format, and read counts covering enhancers (Table S2) extracted using deepTools pyBigWig 54 Palantir data were subsampled to 40 cells/type. northstar's subsample method 55 was used to infer cell states within ME-1 guided by Palantir data 6 . For graph construction, 10 external (non-mutual) neighbours were allowed to compensate for the fact that ME-1 cells are quite distant from actual hematopoietic cells.…”

“…We reasoned that a more sophisticated feature selection together with soft guidance from healthy marrow data could reveal additional hidden heterogeneity. We therefore switched from unsupervised clustering to northstar, a semi-supervised clustering algorithm that leverages information from training data to channel the axes of heterogeneity during feature selection, graph construction, and cell community detection 55 . Using healthy marrow transcriptomes 6 (Figure 1A) as training data, this analysis revealed two major subpopulations, HSC-like (pink) and Mono-precursor-like (purple, 1136 and 277 out of 1489 cells respectively) plus a minor population that was more similar to Ery-precursor cells (lime, 47 out of 1489 cells) and two small groups of cells resembling Megakaryocytes (18 cells) and Monocytes (11 cells, Figure 4D).…”

“…A similar circuit, comprised of GATA2, TAL1, and FLI1 (an ETS TF closely related to ERG) has been previously reported during embryonic HSC specification 72 , while GATA1, TAL1 and KLF1 form a sub-circuit in erythroid cells 73 . Indeed, recycling of regulatory modules is a key feature of developmental networks 38 , underlining the utility of cell classification strategies such as northstar 55 .…”

A comparison of data integration methods for single-cell RNA sequencing of cancer samples

“…Previous single cell batch correction benchmarking studies evaluated algorithm performance on simulated data or datasets derived from healthy tissues and peripheral blood mononuclear cells [4][5][6] . Despite an abundance of data, no single cell batch-effect correction methods are designed for or benchmarked on single cell datasets containing malignant cells 7 . Due to the inherent biological complexity both within and between tumours, these samples present unique technical challenges for batch correction that are not represented in previous benchmarking efforts.…”

Insitutype: likelihood-based cell typing for single cell spatial transcriptomics

“…In tissues with cell types characterized by high-quality reference datasets (Regev 2017, Quake 2021, Allen Institute for Brain Science 2011), cell typing can employ supervised classification, avoiding the instability and interpretation difficulties of clustering. A hybrid approach, semi-supervised cell typing, detects known cell types while also discovering new clusters (Zanini 2020).…”